Iterated Function Systems with OCaml

opam package

Fractals are images of infinite complexity, characterized by being "similar" to themselves in some sense at all scales of magnification.

Iterated function systems (IFS) are a method of generating fractals using self-similarity. An IFS image is defined as being the sum of geometric transforms of itself: each transform is a small affine map (a combination of scaling, rotation and translation) that maps the whole image onto one of its parts. It turns out that simply specifying the transforms along with a weight for each transform is enough to determine the image.

How it works

The program draws the image with the so-called chaos game:

  1. Start from an arbitrary point.
  2. Pick one of the transforms at random, according to its weight.
  3. Apply it to the current point and plot the result.
  4. Repeat from step 2, a few hundred thousand times.

Whatever the starting point, the plotted points quickly converge onto the attractor of the system — the fractal image. More iterations simply fill in the picture with more detail.

In the code (ifs_fractals.ml), a fractal is described by an ifs record:

Available fractals

Sixteen fractals are predefined:

How to use this code

The fractals are drawn in a 400x640 window of the OCaml graphics library, which requires a graphical display — or rendered straight to a PNG file with the -o option (or the save_png function), which requires no display at all.

Installation with opam

The project is packaged as ifs-fractals for opam, the OCaml package manager:

$ opam install ifs-fractals

This installs the ifs-fractals command:

$ ifs-fractals barnsley           # draw a fractal in a window
$ ifs-fractals --list             # list the sixteen available fractals
$ ifs-fractals -n 1000000 lace    # override the number of plotted points
$ ifs-fractals -o fern.png barnsley            # render to a PNG file instead
$ ifs-fractals -o big.png -s 800x1280 maple    # ... at a custom size
$ ifs-fractals -o fern.png -c barnsley         # ... colored by point density

Rendering with -o does not open a window and works without a display, so it can run on a headless server. The PNG writer is built into the library — no image library is needed.

With -c, instead of plotting every point black, each pixel is colored by how often the chaos game visited it, on a log scale from light green (rarely) to dark blue (constantly). This reveals a structure that the flat rendering hides: the attractor is not uniformly dense, and the regions the transforms visit most — stems, midribs, spiral tips — stand out darkest. Density rendering benefits from more points than the defaults, e.g. -n 1000000:

Barnsley Fern colored by density

In your OCaml interpreter

The package also provides a library, so you can play from the interactive toplevel:

$ ocaml
# #use "topfind";;
# #require "ifs-fractals";;
# open Ifs_fractals;;
# draw barnsley 200000;;

draw takes a fractal and the number of points to plot — try smaller values like 20000 to watch the image build up, or replace barnsley with any of the predefined fractals above. To write a PNG file instead of drawing in a window, use save_png (with optional ~width and ~height, defaulting to the window's 400x640):

# save_png barnsley 200000 "fern.png";;
# save_png ~width:800 ~height:1280 maple 200000 "maple.png";;
# save_png ~color:true barnsley 1000000 "fern-density.png";;

The result:

Barnsley Fern

From a clone of the repository

The historical workflow (this project started in 2010 as a single toplevel script) still works without installing the package. Install the dependencies once — on Debian/Ubuntu:

$ sudo apt install ocaml opam
$ opam init
$ opam install graphics ocamlfind
$ eval $(opam env)

then, from the repository root:

$ ocaml
# #use "ifs_fractals.ml";;
# draw barnsley 200000;;

You can also build and run the executable with dune:

$ dune exec -- ifs-fractals lace

Defining your own fractal

Add a new ifs record in lib/ifs_fractals.ml following the same pattern as the predefined ones: choose your affine transforms, give them cumulative probabilities ending at 1.0, and pick po/sz so that the attractor fits in the displayed region. Transform coefficients for many classic fractals can be found in the resources below.

A challenge: the vegvĂ­sir

The vegvisir fractal was born from a challenge: could an IFS draw something like the vegvísir, the Icelandic "wayfinder" stave (famously tattooed on Björk's arm)?

Strictly speaking, no. An IFS attractor is a single self-similar set — the image is a union of shrunken copies of itself — while the real vegvísir has a different rune at the end of each of its eight arms. That asymmetry is simply out of reach for an IFS.

What is reachable is a symmetric idealization, built with two tricks:

The result is not a vegvísir — more an eight-armed rune compass whose every arm ends in an infinitely recursive copy of the whole — but the family resemblance is there:

VegvĂ­sir

Because the symmetry map eats more than half of the random picks without plotting anything new, this fractal needs more iterations than the others to fill in — draw vegvisir 500000;; is a good start.

Fractals from nature

The vegvísir experiment taught a lesson: designed symbols resist IFS, while things that grow embrace it. This is no accident. A plant does not follow a blueprint of its final shape; it grows by repeating simple local rules — sprout, shrink, turn, repeat. Its final form is the accumulation of the same rule applied at every scale, which is exactly what an IFS attractor is: the fern is not like a fractal, it is the fixed point of a handful of affine maps, and so, in a very real sense, are the plants themselves. That is why the most convincing images in this collection — the fern, the maple leaf, the tree — take so few numbers to describe, and why Barnsley needed only four transforms and twenty-four coefficients to capture a fern.

Two more fractals push this idea further.

sunflower is phyllotaxis distilled to two transforms. The first rotates by the golden angle — 137.508°, the angle real plants use to place successive seeds and florets — while contracting slightly toward the center. Because the golden angle is the "most irrational" angle, consecutive buds never line up into spokes; they fill the disk evenly, exactly as in a real sunflower head or a romanesco. The second transform plants a bud at the rim, and since every bud is a copy of the whole attractor, each one is a rosette made of rosettes:

Sunflower

lace is Queen Anne's lace (wild carrot), whose flower is an umbel: a dome of stalks radiating from one point, each stalk ending in a smaller umbel, each of those in smaller umbels still — an umbrella made of umbrellas. Five transforms place shrunken, slightly rotated copies of the entire plant along the rim of the dome, and a sixth squashes the whole image into the thin stem:

Queen Anne's lace

Draw them with draw sunflower 500000;; and draw lace 300000;;.

Paper

The design techniques behind the fractals added in 2026 (stem maps, spiral generators, radial replication, golden-angle maps and non-contracting symmetry maps) are written up in a short paper, Building New IFS Attractors: a Working Vocabulary of Affine Maps.

Resources

This work was carried out during a functional programming course.

Some information about Iterated Function Systems (with the Barnsley Fern):

Alternatives